无保护左转环境下自动驾驶汽车交互策略分析  

Analysis of Autonomous Vehicles'Interaction Strategy in Unprotected Turn Scenarios

作　　者：倪颖[1] 齐骁 杭鹏 周东浩 孙剑[1] NI Ying;QI Xiao;HANG Peng;ZHOU Dong-hao;SUN Jian(Key Laboratory of Road and Traffic Engineering of the Ministry of Education,Tongji University,Shanghai 201804,China)

机构地区：[1]同济大学道路与交通工程教育部重点实验室,上海201804

出　　处：《中国公路学报》2023年第9期271-287,共17页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(52125208,52232015,52272313)。

摘　　要：在人类驾驶汽车(HV)与自动驾驶汽车(AV)构成的混驾环境下,AV交互策略是否符合人类驾驶的预期成为影响新型混合交通流运行安全的重要变量。为了综合、系统地分析AV交互策略特征,基于HV和AV混行驾驶数据集Argoverse2,提取了近4 300组左转和直行冲突交互事件。进一步由表及里,分别从交互即时运动状态、交互过程运动行为和交互内在决策选择3个层面,提出了“冲突点-轨迹线-事件集”一体化三维分析框架。在冲突点维度,从交互时间层面定义了到达冲突点时间差、从动作空间层面定义了到达冲突点协作加速度,并展开分析;在轨迹线维度,利用Frenet坐标转换,选取时间-纵向位移轨迹、横向位移-纵向位移轨迹从纵向行进、横向偏移2个层面分析交互过程;在事件集维度,基于决策树挖掘HV潜在交互准则并与AV交互策略对比分析。结果表明,Argoverse AV左转交互策略与HV存在明显差异:交互即时运动状态表现出更加明细的保守性特征,动作空间与HV相比收缩了31.2%;交互过程行为则表现出策略的单一性特征,停车让行等待位置及通过冲突点行为的选择较HV更为单一,总体横向偏移程度比HV小57.1%;交互内在决策选择与人类潜在交互准则相比存在非社会性特征,AV无法理解HV表达的交互意图,本应先行但却选择让行的交互事件占比高达93.3%。总体而言,Argoverse AV交互特征具有策略保守、行为单一、决策社会性差(理解HV意图能力弱)等显著特征。研究结论可为AV规划决策算法、冲突消解策略提供改进指导方向。In a driving environment consisting of human-driving vehicles(HV)and autonomous vehicles(AV),ensuring that the interaction strategies of AVs align with human driving expectations is pivotal for the operational safety of emerging mixed traffic flows.This research extracted approximately 4300 conflict interaction events between left-turning and straight-through vehicles from the HV and AV mixed driving dataset,Argoverse2.An integrated three-dimensional analysis framework,representing interaction instantaneous motion state,interaction process motion behavior,and interaction internal decision-making,was introduced,delving from superficial interactions to more intrinsic behaviors.This framework encompasses the conflict-point dimension,trajectory-line dimension,and event-set dimension.Within the conflict-point dimension,time to conflict point differences were utilized to evaluate interaction timings,while cooperative acceleration examined the characteristics at the action space level.In the trajectory dimension,Frenet coordinate conversion facilitated the selection of time-longitudinal trajectories and horizontal-longitudinal trajectories,enabling a detailed analysis of vertical processes and horizontal deviations.Within the event-set dimension,potential interaction criteria for HVs were discerned using decision trees and compared with AVs'strategies.Analysis indicated pronounced differences between Argoverse AVs'left-turning interaction strategies and those of HVs.AVs exhibit more conservative instantaneous motion states,with their action space contracting by 31.2%relative to HVs.Their interaction process motion behaviors are more streamlined,as evidenced by the simplified choices in both waiting positions and transiting through conflict points.Overall,the lateral deviation in AVs is 57.1%less than in HVs.Notably,when compared to potential human interaction criteria,AVs'internal decision-making exhibits an evident lack of sociability,leading to challenges in discerning HV intentions.Specifically,in instances where AVs

关 键 词：交通工程 自动驾驶 交互策略特征分析 无保护左转 人机混驾环境 

分 类 号：U491.2[交通运输工程—交通运输规划与管理]